Socially-Aware Machine Learning: Towards Leveraging the Relationship between Narrative Comprehension and Mentalizing

Abstract

Narratives are the fundamental means by which people organize, understand, and explain the social world. Research suggests that exposure to narratives improves mentalizing, referring to the capacity to forecast and reason about others' mental states. Simultaneously, enhanced mentalizing abilities are closely linked to exhibiting improved narrative processing skills. The purpose of this dissertation is to develop modular computational methods that leverage the relationship between mentalizing and narrative comprehension for understanding specific aspects of social-cognitive processes and seek to advance the research towards imparting social awareness to machines. Our work consists of three main functional modules. First, we present a representation learning approach that computes a social situational embedding of sentence-level social events. Next, we apply the learned social event representation to embed, infer and explain the characters' mental states from the narratives. Finally, we analyze some of the basic elements of narrative structure present in short personal narratives as a means of exemplifying the story understanding capability. Particularly, we investigate the role of characters' cognitive tension captured using our inferred mental representation for automatically detecting the central conflict of a story, i.e., the climax and their resolution.

Unlike most previous work that either uses conventional trait-based models or exploits low-level annotations of short fixed-length stories, we tackle a subset of the data and modeling challenges directed at inferring human motives and emotional reactions. First, we construct a relatively open-ended corpus of personal narratives and commonsense knowledge from social media containing more variations in terms of topical content. Using this weakly annotated corpus, we train deep learning models that compute rich representations of social events capturing aspects of syntactic, semantic, and pragmatic properties and integrate them to generate textual explanations of motives and emotions of characters in the narrative. Empirically, our proposed approaches outperform several baselines in mental state tracking tasks and harness transferability to low-resource regimes and other downstream tasks.

As a final contribution in this dissertation, we demonstrate improved narrative processing skills by computationally predicting key elements of narrative structure in personal narratives. Notably, our studies show that integrating the protagonist's mental state embeddings with linguistic information leads to the enhanced prediction of climax and resolution in narratives. Our data and modeling contributions emphasize the value of exploiting the mutual influence of mentalizing and narrative comprehension, thereby promoting future efforts towards building human-centered AI systems.